Reconstruction, geometry, and analysis code for quantitative missing-wedge electron tomography of precipitates

This deposit supports three companion papers on deep-learning missing-wedge correction forquantitative precipitate volume-fraction (VF) estimation in electron tomography. It contains:(a) raw tilt-series projections and the derived sinograms of four datasets — dataA (+-50 deg,101 views, experimental), dataB (+-90 deg, 180 views, experimental, whose +-45 deg subset isthe limited-angle input), and two simulated phantoms dataC and dataD (181 views over +-90 degwith a 91-view +-45 deg subset; dataD retains its native 363-channel detector grid);(b) acquisition geometry (tilt angles, pixel size, detector calibration);(c) trained model weights for Method A (PFITRE/ADMM), Method B (UsiNet; a joint A/C/D model anda B-specific model), and Method C (s83231);(d) reference volumes — the full-angle reconstruction reference (phi-hat) for dataB-dataD,which is the ground truth used for method comparison, and the phantom construction truth(phi-star) for dataC/dataD, which serves as a protocol self-check only; dataA has no full-angleacquisition and therefore no such reference;(e) the reconstruction code, the shared unified VF-protocol code, and three machine-readabletables (vf_table.csv, table2_methodB_cross_dataset.csv, tableS1_tilt_sweep.csv). All volume fractions follow one unified protocol — the "none"-protocol of the companion papers:global P1-P99.9 intensity normalization, two-level Otsu thresholding (vacuum/sample, thenmatrix/bright-phase) with a sample-pixel (non-vacuum) denominator, reported in three definitionsside by side (vf_count, vf_intensity, vf_sumI); see the repository README and code/shared/vf.py. The auxiliary nnU-Net segmentation code used for the threshold-free cross-check is includedunder code/segmentation/ together with its fold-0 best-trained weights; the manual annotationand training pipeline are available from the authors on request and are not part of thedeposited reconstruction methods. Version 5 corrects the Method A dataB reconstruction to its unified-protocol value(VF 20.9455 %; the deposited MethodA_dataB.npy has been replaced accordingly) and adds a"Reproducibility notes" section to the README: the azimuthal missing-wedge FSC of the companionMaterials Characterization paper (Table 2) is reproduced by code/analysis_paper2/fsc_protocol.py(driving compute_table2_fsc.py), and the Method-C production weights were trained with the fixedseed 20260727. Earlier versions added the full-angle reference volumes for dataB-dataD, the dataAacquisition-geometry metadata, the one-shot and 128-slice Method A reconstruction variants, andthe two machine-readable tables table2_methodB_cross_dataset.csv and tableS1_tilt_sweep.csv.